DeepSeek V4.1-Flash: The Multimodal Model That Retires V4 Pro
DeepSeek just retired its own flagship with the smallest model in its new family. V4.1-Flash is multimodal, costs a fraction of what V4 Pro cost, and by the company's own tests beats it on performance, speed and total runtime. From September 14 there is no way back.
What DeepSeek V4.1-Flash Is and Why It Matters
On September 10, 2026, DeepSeek released V4.1-Flash, the smallest model in its new architecture family and the first with native visual understanding. It ships with open weights and a technical report available on Hugging Face, so it can be audited and deployed on your own infrastructure instead of only being consumed through an API.
The move is unusual: rather than unveiling a bigger model than the last one, the company bets on a smaller, cheaper model that performs better. DeepSeek sums it up as more intelligence at a lower cost.
Native vision and a one-million-token context
The model accepts text and images, keeps the one-million-token context of the previous generation and can generate up to 384,000 tokens in a single response. For anyone running agents that read whole repositories or long documents, that headroom stops being the project's bottleneck.
Read also
Asymmetric Architecture: 552B Parameters, 8B Active
V4.1-Flash is a mixture-of-experts model with 552 billion parameters, but they do not all work at once. Its Causal Encoder-Decoder architecture activates only 8 billion parameters on the input side and 16 billion on the output side. That asymmetric design is what delivers big-model capability at small inference cost.
New pretraining methods and larger-scale reinforcement-learning post-training complete the picture. The headline result is that the model lands ahead of the company's own flagships on benchmarks, V4-Pro included.
Smaller KV cache, cheaper agent loops
In agent loops, a large share of the bill does not come from generating text but from re-reading the same context again and again. V4.1-Flash's KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation. Less memory for the same context translates directly into less cost per repeated task.
API Pricing: What Changes on Your Bill
The model is available on the API under the name deepseek-flash. The new prices, per million tokens, are as follows. Input with a cache hit: $0.003 off-peak and $0.006 at peak. Input with a cache miss: $0.15 and $0.30. Output: $0.60 and $1.20.
Peak hours run from 01:00 to 04:00 and from 06:00 to 10:00 UTC, Monday through Friday; off-peak costs half. If you schedule flexible workloads outside those windows, you cut the bill in half.
The concurrency ceiling also rises: 2,500 simultaneous requests against 500 for V4-Pro.
How to Try It Today: API, Open Weights and Tooling
Switching models is a one-line change: just point at deepseek-flash on the API, which keeps compatibility with both the OpenAI and the Anthropic formats. Legacy names still answer for compatibility, but they now serve the new model.
On the tooling side, WorkBuddy (CodeBuddy included) and OpenCode declare same-day support. And if you would rather not depend on an API, the weights are published and the company says it will work with the open-source community on inference support. If that is your route, this blog already covers how to run local models with Ollama without paying per token.
The V4 Pro Retirement: What Changes on September 14
V4-Flash and V4-Flash-Vision-Exp are retired, and their names now route to V4.1-Flash. The important change lands on September 14 at 04:00 UTC: from that moment, every request aimed at deepseek-v4-pro will be routed to V4.1-Flash and billed at Flash rates, until V4.1-Pro arrives. In other words, V4 Pro enters an orderly retirement.
The practical consequence cuts both ways. Anyone with the old model name hardcoded will see no errors, but a different model answering. And anyone paying Pro prices for tasks Flash handles just as well will start paying less without changing a single line.
Conclusion
V4.1-Flash is a clear play: DeepSeek stops competing with itself and standardizes its lineup on a multimodal, cheap, open-weight model that covers the mass-market use case. If you build agents or LLM pipelines, this is the week to review which model you have pinned and how much you are paying for it. On this blog we keep tracking the model race: you can read our analysis of DeepSeek's previous generation launch to see how far it has come.
